A　deep　belief　network　(DBN)　is　one　of　the　most　effective　ways　to　realize　a　deep　learning　technique,　and　has　been　attracting　more　and　more　attentions　in　nonlinear　system　modeling.　However,　it　can　not　provide　satisfactory　results　in　learning　speed　and　modeling　accuracy,　which　is　mainly　caused　by　gradient　diffusion.　To　address　these　problems　and　promote　its　development　in　cross-models,　we　propose　an　efficient　DBN　with　a　fuzzy　neural　network　(DBFNN)　for　nonlinear　system　modeling.　In　this　novel　framework,　DBN　is　considered　as　a　pre-training　technique　to　realize　fast　weight-initialization　and　to　obtain　a　feature-representation　vector.　An　FNN-based　learning　framework　is　developed　for　supervised　modeling　so　as　to　eliminate　the　gradient　diffusion　issue,　where　its　input　happens　to　be　the　feature-representation　vector.　As　a　novel　cross-model,　DBFNN　combines　the　advantages　of　both　pre-raining　technique　of　DBN　and　an　FNN　model　to　improve　nonlinear　system　modeling　capability.　A　classical　benchmark　problem　is　used　to　demonstrate　its　superiority　over　existing　single-models　in　learning　speed　and　modeling　accuracy.